CN110082136A - Rotary machinery fault diagnosis method based on Retrieval method Support Vector Machines Optimized - Google Patents

Abstract

The invention discloses a rotating machinery fault diagnosis method based on cloud genetic algorithm optimization support vector machine. Firstly, in each fault state of the rotating machinery, working signals are extracted from different sensors, and time-frequency feature extraction and reduction are respectively performed to obtain each The eigenvectors of the working signals are obtained based on these eigenvectors to obtain training samples for each fault state, which are divided into training sample set A and training sample set B. First, the training sample set A is used as the training sample, and cloud genetic algorithm is used to The kernel function and penalty factor of the multi-classification model of the network are pre-optimized for parameters, and then the training sample set B is used to optimize again to obtain a multi-classification model. When the rotating machinery fails, the feature vector is extracted from each sensor and input into the multi-classification model. The model gets a diagnosis. The invention can effectively improve the accuracy and efficiency of fault diagnosis of rotating machinery.

Description

Rotating Machinery Fault Diagnosis Method Based on Cloud Genetic Algorithm Optimization Support Vector Machine

technical field

The invention belongs to the technical field of fault diagnosis of construction machinery systems, and more specifically relates to a method for fault diagnosis of rotating machinery based on cloud genetic algorithm optimization support vector machine.

Background technique

With the modernization of industry and the rapid development of science and technology, rotating machinery equipment, as one of the most widely used equipment in the industry, is increasingly used in the fields of electric power, petrochemical industry, aviation and various military industries. Rotating machinery equipment is constantly developing towards high speed, systematization and automation. The scale of its production system is gradually increasing, and the mechanical structure is becoming more and more complex. Each kind of equipment is interrelated and closely coupled with each other. Higher and higher. In the operation of rotating mechanical equipment, accompanied by many uncertain factors, some equipment will inevitably have some failures. Once a key component of a certain equipment fails, a series of chain reactions will occur, and serious ones will cause the entire equipment to fail. The shutdown of the production line will cause huge economic losses and even casualties. The safety, maintainability and reliability of rotating machinery and equipment have become the focus of current research. Establishing and improving the field of fault diagnosis technology can not only ensure the safe operation of rotating machinery and equipment, but also improve economic benefits, reduce maintenance costs and ensure personnel safety. has a positive meaning.

The traditional fault diagnosis technology is not suitable for the diagnosis of complex rotating machinery structure, and has high requirements for operators. With the development of artificial intelligence, intelligent fault diagnosis technology has developed rapidly. There are three main types of methods: (1) model-based fault diagnosis, (2) signal processing-based fault diagnosis and (3) knowledge-based fault diagnosis. Among them, model-based fault diagnosis requires evaluation of design parameters and working conditions to accurately model the system. For complex rotating mechanical equipment, this diagnosis method is uneconomical; fault diagnosis based on signal processing requires analysis and processing of system acquisition signals. Because model-based fault diagnosis is more economical and has certain reliability, but in some specific environments, there are some uncertain information that will have a certain impact on fault information, so the problem of feature extraction and fault judgment need to be further solved; based on Knowledge-based fault diagnosis Since historical fault knowledge, it is not entirely possible to have high reliability like human experts. At the same time, in today's big data era, in the actual application of engineering, the existing intelligent fault diagnosis technology has problems such as speed to be improved, accuracy rate is low, and high-precision classification cannot be achieved, which need to be studied and solved.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies of the prior art, to provide a method for fault diagnosis of rotating machinery based on cloud genetic algorithm optimization support vector machine, to perform time-frequency analysis and processing on the collected rotating machinery working signals, and to optimize the method through cloud genetic algorithm The multi-classification model based on support vector machine realizes the diagnosis of rotating machinery faults, which can effectively improve the accuracy and efficiency of rotating machinery fault diagnosis.

In order to achieve the above-mentioned purpose of the invention, the present invention is based on the cloud genetic algorithm optimization support vector machine fault diagnosis method for rotating machinery comprising the following steps:

S1: Randomly intercept a number of working signals L _s,n (m) of length M from the working signals of different sensors in the R fault states of the rotating machinery, where s=1,2,...,S, S represents the sensor Quantity, n=1,2,...,N _s , N _s represents the number of working signals from the sth sensor, m=0,1,...,M-1, and M>T, T represents the period of the working signal; Note that the label corresponding to each segment of the working signal L _s,n (m) is Y _s,n , and the label Y _s,n is used to identify the fault state of the rotating machine corresponding to the working signal, Y _s,n =1,...,R;

Time-frequency analysis is performed on each working signal in each fault state to extract its time domain features, frequency domain features and time-frequency domain features. The number of each feature is determined according to actual needs, and these features are used as initial features, The initial features of the working signals of different sensors are reduced, and the number of features obtained by reducing the initial features of the sth sensor is D _s , and finally a total of K features are obtained, of which Write down the eigenvector X _s,n corresponding to each working signal L _s,n (m) =(X _s,n (1),X _s,n (2),…,X _s,n (D _s )), where X _s,n (d) represents the value of the dth feature in the working signal L _s,n (m), d=1,2,...,D _s ; the working signal feature vector of each sensor is carried out according to the fault state label Classify, and obtain the eigenvector set φ _s,r of the working signal of the sth sensor in the rth fault state, r=1,2,...,R.

S2: For each fault state, obtain the training samples of the fault state according to the following method: for the rth fault state, randomly select a feature vector X′ _s from the corresponding S feature vector set φ _s,r , _r , and then get a combined feature vector Z=(X′ _1,r ,X′ _2,r ,…,X′ _S,r )=(z ₁ ,z ₂ ,…,z _K ), z _k represents the kth element of the combined feature vector Z, k=1,2,...,K, the fault state label corresponding to the combined feature vector Z is r, which constitutes a training sample; repeat the above process, for each fault state Obtain several training samples;

All training samples are divided into two parts, one part is used as training sample set A, and the other part is used as training sample set B, and the training samples in each training sample set include all fault states;

S3: Construct a multi-classification model based on the support vector machine network, whose input is the combined feature vector, and the output is the fault status label;

S4: Using the training sample set A as the training sample, the cloud genetic algorithm is used to pre-optimize the parameters of the kernel function and penalty factor of the multi-classification model based on the support vector machine network;

S5: The kernel function and penalty factor obtained by pre-optimizing the parameters in step S4 are used as the initial values of the kernel function and penalty factor of the multi-classification model based on the support vector machine network, and the training sample set B is used as the training sample. The kernel function and penalty factor of the multi-classification model of the support vector machine network are optimized to obtain the final multi-classification model;

S6: When the rotating machinery fails, use S sensors to collect and obtain S working signals with a length of M Extract the K features obtained from the reduction in step S1 Compose the combined eigenvectors Input it into the multi-classification model trained in step S5 to obtain the diagnosis result.

The present invention optimizes the rotating machinery fault diagnosis method based on the cloud genetic algorithm support vector machine. Firstly, in each fault state of the rotating machinery, the working signals are extracted from different sensors, and the time-frequency features are extracted and reduced respectively to obtain the characteristics of each working signal. Based on these eigenvectors, the training samples of each fault state are obtained, which are divided into training sample set A and training sample set B. First, the training sample set A is used as the training sample, and the multi-classification based on the support vector machine network is performed by cloud genetic algorithm. The kernel function and penalty factor of the model are pre-optimized for parameters, and then the training sample set B is used to optimize again to obtain a multi-classification model. When the rotating machinery fails, the feature vector is extracted from each sensor, and the multi-classification model is input to obtain the diagnosis result. .

The present invention has the following beneficial effects:

1) The present invention has carried out comprehensive time-frequency feature extraction to the original working signal, avoids the feature imperfection of the signal, and screens redundant features and invalid features at the same time, so as to obtain feature information exactly associated with the fault state;

2) The present invention improves the genetic algorithm, adopts the cloud genetic algorithm to pre-optimize and optimize the parameters of the multi-classification model based on the support vector machine network, effectively reduces redundant searches, and makes the population evolution towards the optimal direction This reduces the running time of diagnosis and improves the accuracy of network diagnosis;

3) The multi-classification model based on the support vector machine network proposed by the present invention has the advantages of simple structure and few parameters, low requirements on hardware resources, low economic requirements, and certain generalization ability, and it is proved by experiments that the present invention can Effectively improve the accuracy of fault diagnosis of rotating machinery.

Description of drawings

Fig. 1 is the specific embodiment flowchart of the fault diagnosis method based on cloud genetic algorithm optimization support vector machine of the present invention;

Fig. 2 is the flow chart that adopts cloud genetic algorithm to carry out parameter optimization to the multiclassification model based on support vector machine network among the present invention;

Fig. 3 is that the present invention and comparative method in the present embodiment are for the fault diagnosis accuracy statistical figure of U.S. Case Western Reserve University data;

Fig. 4 is the time-consuming comparison chart of fault diagnosis of U.S. Case Western Reserve University data between the present invention and the comparative method in the present embodiment;

Fig. 5 is the statistical diagram of the accuracy rate of fault diagnosis of the autonomous experiment platform data between the present invention and the comparison method in the present embodiment;

Fig. 6 is a time-consuming comparison chart of the fault diagnosis of the autonomous experiment platform data between the present invention and the comparison method in this embodiment.

Detailed ways

Specific embodiments of the present invention will be described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

FIG. 1 is a flow chart of a specific embodiment of the method for diagnosing a rotating machinery fault based on cloud genetic algorithm optimization support vector machine of the present invention. As shown in Figure 1, the specific steps of the rotating machinery fault diagnosis method based on cloud genetic algorithm optimization support vector machine of the present invention include:

S101: Working signal data processing:

Randomly intercept a number of working signals L _s,n (m) of length M from the working signals of different sensors in the R fault states of the rotating machinery, where s=1,2,...,S, S represents the number of sensors, n=1,2,...,N _s , N _s represents the number of working signals from the sth sensor, m=0,1,...,M-1, and M>T, T represents the period of the working signal; The label corresponding to the segment working signal L _s,n (m) is Y _s,n , and the label Y _s,n is used to identify the fault state of the rotating machine corresponding to the working signal, Y _s,n =1,...,R.

Time-frequency analysis is performed on each working signal in each fault state to extract its time domain features, frequency domain features and time-frequency domain features. The number of each feature is determined according to actual needs, and these features are used as initial features, The initial features of the working signals of different sensors are reduced, and the number of features obtained by reducing the initial features of the sth sensor is D _s , and finally a total of K features are obtained, of which Write down the eigenvector X _s,n corresponding to each working signal L _s,n (m) =(X _s,n (1),X _s,n (2),…,X _s,n (D _s )), where X _s,n (d) represents the value of the dth feature in the working signal L _s,n (m), d=1,2,...,D _s . Classify the working signal eigenvectors of each sensor according to the fault state label, and obtain the eigenvector set φ _s,r of the working signal of the sth sensor in the rth fault state, r=1,2,...,R.

S102: Obtain training samples:

For each fault state, the training samples of the fault state are obtained according to the following method: for the rth fault state, a eigenvector X′ _s,r is randomly selected from the corresponding S eigenvector set φ _s,r , respectively, Then according to the combination of the sensor numbers, a combination feature vector Z=(X′ _1,r ,X′ _2,r ,…,X′ _S,r )=(z ₁ ,z2 _, …,z _K ), z _k represents the combination The kth element of the feature vector Z, k=1, 2, ..., K, the fault state label corresponding to the combined feature vector Z is r, which constitutes a training sample. Repeat the above process to obtain several training samples for each fault state.

Divide all training samples into two parts, one part is training sample set A, and the other part is training sample set B, and the training samples in each training sample set contain all fault states. The present invention constructs two training sample sets to be convenient to follow-up parameter optimization work, can divide randomly when dividing training sample set, but because classification model in the present invention adopts the classification model based on support vector machine network, and in support vector machine network, The boundary vector is very important for the training of the support vector machine network, so try to place the vector near the boundary in the training sample set A. In order to improve the efficiency of subsequent parameter optimization, a training sample set division method is proposed in this embodiment:

For each fault state, calculate the average vector of all its training samples, then calculate the distance between each training sample and the average vector, sort the training samples according to the distance from large to small, and select the first Q training samples to join the training sample set A, where the value of Q is determined according to actual needs, and the rest of the training samples are added to the training sample set B.

S103: Build a multi-classification model:

A multi-classification model based on a support vector machine network is constructed, whose input is a combined feature vector and the output is a fault state label.

In the present invention, there are R kinds of fault types of the rotating machinery, and there are K characteristic parameters related to the fault state of the rotating machinery, and a multi-classification model based on a support vector machine network is constructed to classify and identify various fault characteristics of the rotating machinery. In this embodiment, the RBF kernel function is adopted, and a one-to-one method is selected to construct a multi-classification model, wherein the number of nodes in the input layer is K, and the number of nodes in the output layer is R.

S104: Pre-optimization of multi-classification model parameters:

Taking the training sample set A as the training sample, the cloud genetic algorithm is used to pre-optimize the parameters of the kernel function and penalty factor of the multi-classification model based on the support vector machine network.

The Cloud Genetic Algorithm has the advantages of the Genetic Algorithm, and at the same time, the Genetic Algorithm has been improved to make the overall evolution faster, improve the convergence speed, and reduce unnecessary searches. Fig. 2 is a flow chart of parameter optimization of multi-classification model based on support vector machine network using cloud genetic algorithm in the present invention. As shown in Figure 2, in the present invention, adopting cloud genetic algorithm to carry out the concrete steps of parameter optimization to the multi-classification model based on support vector machine network comprises:

S201: Real number encoding:

Real-number encoding of the kernel function and penalty factor of the multi-classification model based on the support vector machine network. Traditional genetic algorithms are binary coded. As the complexity of the problem increases and the accuracy improves, the coding length of the simple genetic algorithm will also increase, occupying a large amount of memory and greatly reducing the performance of the algorithm. The present invention adopts the way of real number coding to improve the algorithm performance.

S202: Initialize the population:

The kernel function and the real code of the penalty factor are used as individuals in the cloud genetic algorithm population, and the population is initialized, and the population size is recorded as Q.

S203: Calculate individual fitness:

For each individual in the population, use its corresponding kernel function and penalty factor to set the multi-classification model based on the support vector machine network, use the training samples to cross-validate the multi-classification model, and use the obtained classification accuracy as the individual adaptation Obviously, the greater the classification accuracy, the better the individual.

S204: Determine whether the termination iteration condition is satisfied, if so, go to step S205, otherwise go to step S206. For the fault diagnosis of rotating machinery, there are two conditions for terminating iterations, one is whether the maximum evolution algebra is reached; the other is whether the accuracy of fault diagnosis does not change significantly within a certain period of time, which can be selected according to the needs in practical applications.

S205: Get the parameter optimization result:

Select the optimal individual from the current population, and use its corresponding kernel function and penalty factor as the parameter optimization result.

S206: Select operation:

Select individuals in the population. The selection operation is completed using the tournament selection mechanism to ensure that individuals with higher individual fitness values are retained for the next step of genetic operations, thereby accelerating the evolution of the entire population.

S207: Cross operation:

The traditional genetic algorithm is to randomly pair the individuals in the population, and exchange some chromosomes according to the crossover probability of the individual. However, this approach makes the evolutionary direction of the genetic algorithm unknown and uncontrollable. The cloud model is an uncertainty model, which can convert between qualitative concepts and quantitative values based on traditional probability and statistics theory and fuzzy theory. It uses three numerical characteristics of cloud expectation, entropy and hyperentropy to represent the numerical characteristics of qualitative concepts and their uncertainty. The Y-conditional cloud generator is a cloud generator under the condition of given the above three numerical characteristics and a given degree of certainty. In order to reduce unnecessary searches and speed up the evolution, the crossover operation uses the Y conditional cloud model instead of the crossover probability.

The specific method of the crossover operation in the present invention is: first, obtain the certainty μ of each parent individual selected in step S206 through linear function calculation; then calculate the digital characteristics of the cloud model, including cloud expectation _{Ex, entropy E n and} super Entropy H _e ; according to the calculated degree of certainty and three digital characteristics of the cloud model, execute the Y conditional cloud generator to obtain a pair of offspring, and calculate the individual fitness of the offspring obtained and each parent individual selected in step S206 , select better Q individuals.

The calculation formula of the degree of certainty μ is as follows:

Among them, F _max and F _min represent the global "maximum" and "minimum" fitness values of the contemporary population respectively, F' is the greater of the fitness of the crossed two parents, μ _max and μ _min are the values of the artificially specified degree of certainty Maximum and minimum values.

The calculation formulas of cloud expectation E _x , entropy E _n and hyper-entropy He _e are as follows:

Among them, E _x represents the cloud expectation, F _f and F _m represent the fitness of the two parents respectively, X _f and X _m represent the two parent individuals performing the crossover operation, E _n represents the entropy, and x _fmax represents the current population The individual corresponding to the highest fitness, x _fmin represents the individual corresponding to the lowest fitness of the current population, He represents the _hyper -entropy, and c ₁ and c ₂ are constant parameters specified by humans.

It can be seen from the expression of the cloud expectation Ex that choosing the linear function of the two parents to express the expectation _{Ex is} beneficial for the next generation to be closer to the one with the higher individual fitness value. The entropy E _n is proportional to the search area, the greater the entropy, the larger the coverage of the cloud, and thus the larger the search area of the individual in the crossover operation. In the Y conditional cloud generator, the normal cloud is a kind of pan-normal distribution. According to the "3σ" principle of the normal distribution, combined with the accuracy and speed of the evolutionary algorithm, usually choose 6≤c ₁ ≤3p (p is the population size ). At the same time, as the _hyperentropy He increases, the stable trend will decrease to some extent, but if the _hyperentropy He decreases, randomness will be lost. To sum up, the parameters should be reasonably selected according to the actual situation.

S208: mutation operation:

Perform mutation operation on the Q individuals obtained in step S207, and return to step S203.

S105: Multi-classification model parameters are optimized again:

The kernel function and penalty factor obtained by pre-optimizing the parameters in step S3 are used as the initial values of the kernel function and penalty factor of the multi-classification model based on the support vector machine network, and the training sample set B is used as the training sample, and the cloud genetic algorithm is used to analyze the The kernel function and penalty factor of the multi-classification model of machine network are optimized to obtain the final multi-classification model.

S106: Fault diagnosis:

When the rotating machinery fails, use S sensors to collect and obtain S working signals with a length of M Extract the K features obtained from the reduction in step S1 Compose the combined eigenvectors Input it into the multi-classification model trained in step S105 to obtain the diagnosis result.

Example

In order to better illustrate the technical scheme and technical effect of the present invention, a specific example is used to analyze and illustrate the working process and technical effect of the present invention. Bearing failure is a typical failure in rotating machinery. Therefore, in this embodiment, the open data of bearing failure of Case Western Reserve University in the United States and the bearing failure data collected by a self-built fault simulation platform are used for experimental testing.

For the bearing fault data of Case Western Reserve University in the United States, six types of faults were selected, namely (1) slight damage to the rolling element (B1); (2) severe damage to the rolling element (B2); (3) slight damage to the inner ring (I1) ;(4) Serious damage to the inner ring (I2); (5) Slight damage to the outer ring (O1); (6) Serious damage to the outer ring (O2). The acceleration vibration signal collected by the sensor and the acceleration sensor at the base end. Through random selection, a total of 100 sets of samples are selected for each fault type, and each set of data contains 1024 sample points. Table 1 is the bearing fault information table constructed based on the bearing fault data of Case Western Reserve University in the present embodiment.

Table 1

For the self-built fault simulation platform, five types of faults are set: (1) slight damage to the inner ring (I1); (2) moderate damage to the inner ring (I2); (3) severe damage to the inner ring (I3); (4) Moderate damage to the outer ring (O2); (5) slight damage to the outer ring (O1). For each fault type, two acceleration sensors HD-YD-221 located in the vertical direction and horizontal direction of the bearing housing are randomly selected to collect acceleration vibration signals (A0, A1) and the model is WT0180, which is used to measure the vertical and horizontal of the shaft There are 200 groups of 3 displacement sensors (V0-V2) for the direction voltage information, and each group contains 1024 sample points. Table 2 is the bearing fault database collected and constructed based on the self-built fault simulation platform in this embodiment.

Table 2

In this embodiment, time-frequency analysis and processing are performed on each segment of the working signal. The time-domain characteristics of the working signal in this embodiment are: root mean square value, root square amplitude, absolute mean square amplitude, kurtosis factor, form factor, kurtosis, skewness, pulse factor, peak value, margin The coefficients and frequency domain characteristics are: mean square frequency, center of gravity frequency, and frequency variance. Then, the wavelet packet analysis method is used to analyze the time-frequency characteristics of the working signal, and a total of 8 time-frequency domain parameters from E1 to E8 are calculated, which are used as time-frequency domain parameters. Features, where E1-E8 is a time-frequency domain characteristic parameter after the energy of the signal is reconstructed and normalized after the wavelet analysis. Therefore, 11 time-domain features, 3 frequency-domain features, and 8 time-frequency domain features can be obtained for each working signal, and a total of 22 features can be obtained. The above features are often used in time-frequency analysis, and their specific calculation formulas will not be repeated here.

Since the present invention is aimed at a rotating machine with multi-sensors, that is, multi-source working signals, features need to be extracted from the working signals of each sensor, so there are a total of S×22 features. For the data of Case Western Reserve University in the United States, there are 3 sensors, and a total of 3*22=66 features are obtained, and then feature reduction is performed on them. In this embodiment, rough set theory is used to perform feature reduction. Table 3 is the characteristic statistical table of the data of Case Western Reserve University in the United States after rough set theory characteristic reduction.

Time-Frequency Features Time-Frequency Features Time-Frequency Features 1 drive side_maximum 11 Driver_E4 twenty one Fan end_E6 2 Drive end_rms value 12 Driver_E5 twenty two Base end_RMS 3 Drive end_square root amplitude 13 Driver_E6 twenty three base end_square root amplitude 4 Driver_Form factor 14 Fan end_RMS twenty four base end_kurtosis 5 Drive end_pulse factor 15 Fan side_Kurtosis 25 Base end_skew 6 drive side_peak 16 Fan side_pulse factor 26 Base end_Center of gravity frequency 7 Drive end_Center of gravity frequency 17 Fan side_margin factor 27 Base end_E1 8 Drive side_frequency variance 18 Fan end_skew 28 Base end_E2 9 Driver_E2 19 Fan side_E1 29 Dock side_E3 10 Driver_E3 20 Fan end_E3 30 Base end_E4 31 Base end_E7 32 Base end_E8

table 3

For the data of the self-built fault platform, there are 5 sensors, and a total of 5*22=110 time-frequency features are obtained. Table 4 is the feature statistics table of the self-built fault platform data after rough set theory feature reduction.

Table 4

The time-frequency characteristics of the obtained two parts of bearing fault data are divided into training sample set and test sample set. For the data of Case Western Reserve University in the United States, 80 sets of data are selected as the training sample set, and 10 sets of data are randomly selected as the training sample set A to pre-optimize the parameters of the multi-classification model, and the remaining 70 sets of data are used as the training sample set B for multi-classification The parameters of the model are optimized again, and another 20 sets of data are selected as the test sample set. Because 6 kinds of fault states of rolling bearings are selected, 32 characteristic parameters related to the fault states of rolling bearings are screened out to provide reference for the diagnosis of rolling bearing fault states. Therefore, the multi-classification model based on the support vector machine network has 32 input layer nodes and 6 output layer nodes.

For the self-built experimental platform data, 180 sets of data were selected as the training sample set, 18 sets of data were randomly selected as the training sample set A to pre-optimize the parameters of the multi-classification model, and the remaining 162 sets of data were used as the training sample set B to perform parameters for the multi-classification model Optimize again, and select another 20 sets of data as the test set. Since 5 kinds of fault states were obtained, 69 characteristic parameters related to the fault state of rolling bearings were screened out to provide reference for the diagnosis of fault states of rolling bearings. Therefore, the number of input layer nodes of the multi-classification model based on support vector machine network is 69, and the number of output layer nodes is 5. Among them, the training sample set is used for training the multi-classification model, and the test sample set is used for testing the multi-classification model to count the classification accuracy. Table 5 is the parameter setting in the cloud genetic algorithm in this embodiment.

μ_max μ_min c₁ c₂ 0.9 0.1 40 10

table 5

In order to better illustrate the technical effects of the present invention, the multi-classification model based on the support vector machine network optimized by the traditional genetic algorithm is used as a comparison method, and the multi-classification model based on the support vector machine network optimized by the cloud genetic algorithm of the present invention Conduct multiple experiments together to count the accuracy of fault diagnosis and the running time of fault diagnosis. Table 6 is a statistical table of the fault diagnosis results of the present invention and the comparative method for the data of Case Western Reserve University in the United States and the data of the independent experiment platform in this embodiment.

Table 6

Fig. 3 is a statistical diagram of the fault diagnosis accuracy rate of the present invention and the comparative method for the data of Case Western Reserve University in the United States in this embodiment. Fig. 4 is a time-consuming comparison chart of fault diagnosis of data from Case Western Reserve University in the United States between the present invention and the comparative method in this embodiment. Fig. 5 is a statistical diagram of the fault diagnosis accuracy rate of the present invention and the comparative method for the autonomous experiment platform data in this embodiment. Fig. 6 is a time-consuming comparison chart of the fault diagnosis of the autonomous experiment platform data between the present invention and the comparison method in this embodiment.

It can be seen from Table 6 and Fig. 3 to Fig. 6 that for the two kinds of fault data, the present invention shows a higher accuracy rate, and at the same time, the accuracy rate has a certain stability in multiple experiments. Compared with the data of Case Western Reserve University in the United States, the number of samples of the independent experiment platform is larger. Under the condition of large data and large samples, the present invention adopts cloud genetic algorithm to optimize the multi-classification model based on support vector machine network to have faster diagnosis performance and has a certain generalization ability, which shows that the multi-classification model proposed by the present invention can achieve a higher diagnosis rate in a shorter time, and is effective in the field of fault diagnosis of rotating machinery.

Although the illustrative specific embodiments of the present invention have been described above, so that those skilled in the art can understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, As long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are included in the protection list.

Claims (3)

1. a kind of rotary machinery fault diagnosis method based on Retrieval method Support Vector Machines Optimized, which is characterized in that including Following steps:

S1: several length are intercepted at random from the working signal of different sensors under the R kind malfunction of rotating machinery respectively For the working signal L of M_s,n(m), wherein s=1,2 ..., S, S indicate number of sensors, n=1,2 ..., N_s, N_sIt indicates to come from s The working signal quantity of a sensor, m=0,1 ..., M-1, and M > T, T indicate the period of working signal；Remember every section of work Signal L_s,n(m) corresponding label is Y_s,n, label Y_s,nFor the malfunction of the corresponding rotating machinery of identification work signal, Y_s,n =1 ..., R；

For every kind of malfunction every section of working signal carry out time frequency analysis processing with extract its temporal signatures, frequency domain character and Time and frequency domain characteristics, the quantity of every kind of feature determine according to actual needs, using these features as initial characteristics, then for difference The initial characteristics of the working signal of sensor carry out reduction, and the feature quantity that s-th of sensor initial characteristics reduction of note obtains is D_s, it is finally total to obtain K feature, whereinRemember each working signal L_s,n(m) corresponding feature vector, X_s,n =(X_s,n(1),X_s,n(2),…,X_s,n(D_s)), wherein X_s,n(d) working signal L is indicated_s,n(m) value of d-th of feature, d=in 1,2,…,D_s；The working signal feature vector of each sensor is classified according to malfunction label, obtains s-th of biography Feature vector set φ of the working signal of sensor under r kind malfunction_s,r, r=1,2 ..., R.

S2: for each malfunction, the training sample of the malfunction is obtained in accordance with the following methods respectively: for the event of r kind Barrier state, from corresponding S feature vector set φ_s,rIt is middle randomly select respectively a feature vector, X '_s,r, then according to sensing Device serial number combines to obtain an assemblage characteristic vector Z=(X '_1,r,X′_2,r,…,X′_S,r)=(z₁,z₂,…,z_K), z_kIndicate combination K-th of element of feature vector Z, k=1,2 ..., K, the corresponding malfunction label of assemblage characteristic vector Z are r, that is, constitute one A training sample；Above procedure is repeated, several training samples are obtained respectively to every kind of malfunction；

All training samples are divided into two parts, a part is used as training sample set A, and another part is as training sample set B, often The training sample that a training sample is concentrated includes all malfunctions；

S3: more disaggregated models of the building based on support vector machines network, input are assemblage characteristic vector, are exported as malfunction Label；

S4: using training sample set A as training sample, using Retrieval method to more classification moulds based on support vector machines network The kernel function and penalty factor of type carry out the processing of parameter pre-optimized；

S5: the kernel function and penalty factor that step S4 parameter pre-optimized is obtained are as more classification based on support vector machines network The kernel function of model and the initial value of penalty factor, using training sample set B as training sample, using Retrieval method to being based on The kernel function and penalty factor of more disaggregated models of support vector machines network carry out parameter optimization, obtain final more classification Model；

S6: when rotating machinery breaks down, S sensor is used to collect S length as the working signal of MFrom In extract K feature, form assemblage characteristic vectorIt is input to trained more points of step S5 In class model, diagnostic result is obtained.

2. rotary machinery fault diagnosis method according to claim 1, which is characterized in that training sample in the step S2 It divides method particularly includes:

For each malfunction, the average vector of its all training sample is calculated, each training sample is then calculated and this is flat Training sample is ranked up by the distance of equal vector from big to small according to distance, and training sample is added in Q training sample before selection Collect A, wherein determine according to actual needs, training sample set A is added in remaining training sample to the value of Q.

3. rotary machinery fault diagnosis method according to claim 1, which is characterized in that described to use Retrieval method pair The kernel function kernel function and penalty factor of more disaggregated models based on support vector machines network carry out the specific steps of parameter optimization Include:

S4.1: kernel function and penalty factor to more disaggregated models based on support vector machines network carry out real coding；

S4.2: using the real coding of kernel function and penalty factor as the individual in Retrieval method population, initialization population, kind Group's size is denoted as Q；

S4.3: each individual in population, using its corresponding kernel function and penalty factor to based on support vector machines network More disaggregated models be configured, cross validations, the nicety of grading that will be obtained are carried out to the more disaggregated models using training sample As individual adaptation degree；

S4.4: judge whether otherwise meeting termination iterated conditional enters step S4.6 if it is satisfied, then entering step S4.5；

S4.5: selecting optimum individual from current population, using its corresponding kernel function and penalty factor as parameter optimization result；

S4.6: the individual in population is selected；

S4.7: the degree of certainty μ of each parent individuality selected in step S4.6 is calculated by linear function；Then cloud is calculated The numerical characteristic of model, including cloud it is expected E_x, entropy E_nWith super entropy H_e；Further according to three of the degree of certainty and cloud model that are calculated Numerical characteristic executes Y condition cloud generator and obtains a pair of of offspring, calculates and each of selects in obtained offspring and step S4.6 The individual adaptation degree of parent individuality selects preferably Q individual；

S4.8: mutation operation, return step S4.3 are carried out to the Q individual that step S4.7 is obtained.